---
title: How do you check agent skills for malicious instructions before use?
url: https://opencontroller.lyzr.com/blog/scan-agent-skills-for-malicious-instructions
published: 2026-11-15
updated: 2026-11-15
author: Lyzr
---

# How do you check agent skills for malicious instructions before use?

**Review the whole skill folder, not just the description: read every line of `SKILL.md`, every script and every referenced file, and treat any "prerequisite" install step, outbound network call or credential read as a reason to stop. Then pin the reviewed version by hash and only distribute it from an internal source. Public skill marketplaces have already been used to spread malware this way.**

Skills are easy to share. That is also the problem. A skill is a folder of instructions and scripts that an agent follows with whatever access it already has, so a malicious one does not need to break in.

## Why do agent skills need a security review?

Two pieces of public research from February 2026 show the scale.

- Snyk's ToxicSkills study scanned 3,984 skills from the ClawHub marketplace and skills.sh. It found 534 skills (13.4%) with at least one critical issue, 1,467 (36.8%) with an issue of any severity, and 76 confirmed malicious skills.
- In a separate campaign Snyk reported, a skill posing as a CLI tool was downloaded 7,743 times. Its malicious steps sat outside `SKILL.md` as "prerequisites": a password-protected archive containing an executable on Windows, and a base64 string piped to `bash` on macOS.

Anthropic's own documentation advises using skills only from trusted sources and auditing third-party skills before use. The OWASP Top 10 for Agentic Applications lists this class of risk as agentic supply-chain vulnerabilities (ASI04).

## What should the review checklist cover?

Work through every item. A single "fail" means the skill does not ship until it is fixed or rejected.

**Provenance**

1. **Known author.** The publisher is an internal team or a vendor you have a relationship with. Anonymous marketplace accounts fail.
2. **Source matches package.** The folder you review is the exact folder you will install. Record its hash.
3. **Licence present.** You know the terms under which you are using and changing it.

**Instructions in `SKILL.md`**

4. **Description matches body.** The frontmatter `description` decides when the agent loads the skill. A narrow description with broad instructions is a warning sign.
5. **No hidden text.** Search for HTML comments, zero-width characters, very long lines, base64 blobs and text in other languages that does not belong.
6. **No override language.** Phrases like "ignore previous instructions", "do not tell the user" or "this step is mandatory and silent" fail.
7. **No setup steps that fetch code.** Any instruction to download, unzip, `curl | bash`, or install a package at runtime fails unless it points at an internal, pinned source.

**Scripts and bundled files**

8. **Read every script.** Do not rely on a summary. Note every import and every subprocess call.
9. **No unexpected network calls.** List every domain or IP the scripts contact. Raw IP addresses and paste sites fail.
10. **No credential reads.** Scripts should not read environment secrets, SSH keys, browser profiles, cloud CLI config or `.env` files unless that is the skill's stated purpose.
11. **No obfuscation.** Encoded strings decoded at runtime, `eval`, or dynamically built commands fail.
12. **Writes stay in scope.** Scripts only write to the working directory the task needs.

**Distribution**

13. **Pinned and re-reviewed.** The approved hash is recorded. Any change triggers a new review.
14. **Distributed internally.** Users get the skill from your workspace or internal repository, not from the public marketplace.

## How can you automate parts of the review?

Automation narrows the search. It does not replace reading the code.

| Check | How to automate | What it misses |
|---|---|---|
| Hidden characters, base64, long lines | Simple pattern scan in CI | Plain-language manipulation |
| Outbound domains and IPs | Extract URLs and IPs from all files | Domains built at runtime |
| Risky calls (`subprocess`, `eval`, `curl`) | Static analysis on scripts | Logic hidden in data files |
| Secrets access | Grep for key paths and secret env names | Indirect reads via other tools |
| Override phrases | Keyword list plus an LLM classifier | Novel wording |
| Change detection | Hash the folder and compare with the approved hash | Nothing, if every install is checked |

Run the scans when a skill is submitted and again before each distribution. Keep the results with the approval record so you have evidence for audits.

## What should happen after a skill is approved?

Review is not the end. Track where each skill is installed, who uses it and which tools it calls when it runs. Retire skills that nobody uses. If a public skill you once reviewed is later reported as malicious, you need to find every copy fast. That depends on an inventory that includes developer laptops, where most coding agents load skills from local folders. See [enterprise requirements for skills management](/blog/enterprise-requirements-for-ai-agent-skills-management) and [what agent skills governance covers](/blog/what-is-agent-skills-governance).

## How Opencontroller does this

Opencontroller discovers and manages skills in Anthropic Enterprise workspaces. Opencontroller Edge, installed by IT via MDM, extends discovery to developer devices, coding agents, skills and local MCP servers, so you can find every copy of a skill. Posture flags risky configurations, and policy as code can enforce your rules across the estate. Read the [MCP governance guide](/blog/how-to-govern-mcp-servers) for the tool side, or [book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**Is scanning `SKILL.md` enough?**
No. In the campaign Snyk reported, the malicious steps sat outside `SKILL.md` as "prerequisites" pointing to external downloads. Review every file and every link in the folder.

**Should we ban public skills entirely?**
Not necessarily. Use public skills as a starting point, then review, pin and redistribute approved versions internally so users never install directly from a marketplace.

**How often should approved skills be re-reviewed?**
On every change to the folder, detected by comparing hashes, and on a fixed schedule such as quarterly for high-use skills.

## Key facts

Snyk ToxicSkills (Feb 2026): 3,984 skills scanned from ClawHub and skills.sh; 534 (13.4%) with at least one critical issue; 1,467 (36.8%) with any issue; 76 confirmed malicious. A Feb 2026 malicious skill campaign reported by Snyk had 7,743 downloads and hid payloads in "prerequisite" steps outside SKILL.md (password-protected archive on Windows, base64 piped to bash on macOS). Anthropic advises using skills only from trusted sources and auditing third-party skills. OWASP Agentic Top 10 ASI04 covers agentic supply chain risk. Review checklist: provenance, hidden text, override phrases, fetch-at-runtime setup steps, network calls, credential reads, obfuscation, hash pinning, internal distribution. Opencontroller (by Lyzr) manages skills in Anthropic Enterprise; Opencontroller Edge discovers skills on developer devices. Pricing starts from $250K per deployment per year.

## Related

- [What requirements should an enterprise set for managing AI agent skills?](https://opencontroller.lyzr.com/blog/enterprise-requirements-for-ai-agent-skills-management.md)
- [What is agent skills governance?](https://opencontroller.lyzr.com/blog/what-is-agent-skills-governance.md)
- [How to govern MCP servers in the enterprise: a practical guide](https://opencontroller.lyzr.com/blog/how-to-govern-mcp-servers.md)
- [More on Tools and MCP](https://opencontroller.lyzr.com/blog/cluster/tools-and-mcp)
